Every Transformer Has a Story: How Hindsight Memory Helps Our Agent Read It

Prajwal Bevoor — Sat, 07 Nov 2026 17:39:19 +0000

Every electrical transformer has its own operating history.

Two transformers can have the same rating and design, yet live completely different lives. One runs under steady load for years. Another faces frequent overloads. One has a simple maintenance record, while another has been opened up and repaired several times.

These differences matter when we judge transformer health.

A condition report tells us about a transformer at one moment in time. But to truly understand its condition, we need more than the latest report. Previous test results, loading events, maintenance work and inspection notes all explain how that condition developed.

That's the idea behind our project, Transformer Health Historian: a DGA trend agent with memory. We believe transformer monitoring should understand the complete history of each transformer, not just individual measurements.

A Transformer Is More Than Its Latest Test Report

Imagine an engineer receiving a new DGA (dissolved gas analysis) report.

It lists hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide and carbon dioxide. The engineer compares the values with diagnostic guidance.

But the report alone doesn't say whether those gases have increased, decreased or stayed stable. For that, the engineer has to dig out earlier reports, check dates and find maintenance records, and do it again for every transformer in the fleet.

The problem isn't a lack of information. In most cases, the information already exists. It's just not available in one connected view.

Same Number, Different Story

General reference limits are important, but they don't describe the normal behaviour of each individual transformer.

Consider two transformers:

Transformer A has shown acetylene at 3 ppm in several reports in a row.
Transformer B has always shown 0 ppm, and now shows 3 ppm.

Both have the same latest value, but completely different histories.

This doesn't mean A is automatically healthy or that B has a confirmed fault. It means the same number can carry a different meaning depending on its past. Engineers need to look at both the absolute value and the change over time.

Why Loading Events Should Be Remembered

Loading is a big part of a transformer's history. Demand changes through the day, across seasons and with the needs of the network. Higher loading raises operating temperature, depending on design and cooling.

Here's a real example from our system. Transformer T-2 ran at 112–118% load for about three weeks during a feeder transfer, and its top oil temperature reached 92 °C. The next oil test showed carbon monoxide and carbon dioxide clearly up, which can point to paper insulation stress.

Timing alone doesn't prove a cause. But when the agent remembers that overload, it can put the two events side by side, so the engineer starts the investigation with the right context instead of a blank page.

Why Maintenance Records Matter Just as Much

Transformers go through inspections, oil treatment, repairs and tap-changer servicing over their life. When an unusual reading appears after maintenance, engineers naturally want to review that work.

Take transformer T-3:

Its earlier DGA reports showed acetylene at 0 ppm.
Its tap changer was serviced in September.
The next report showed 4 ppm.

An isolated check focuses only on the latest number. A history-aware agent also notices the maintenance that happened just before. It doesn't claim the maintenance caused the change. It highlights the timing and suggests reviewing those records.

Learning From Previous Events

One of the most valuable things about keeping history is learning from earlier abnormal events.

Over time, an organisation builds up records of past issues and failures. If a transformer starts showing a pattern similar to an earlier case, engineers can look at what happened back then.

It doesn't guarantee the same outcome, since every transformer responds differently. But history helps flag the observations that deserve a closer look.

How We Gave the Agent a Memory

To make this work, we used Hindsight, an open-source agent memory system by Vectorize. It lets our agent store events and recall them when they matter.

Our agent connects three kinds of history for every transformer:

1. DGA history: past and current gas readings with test dates.

2. Operating history: loading events and other operating conditions.

3. Maintenance history: what work was done, when, by whom, and what was found.

Every new event is retained in memory. When an engineer asks a question, the agent recalls that transformer's history plus lessons from the rest of the fleet:

memories = self.recall(f"{tx_id} {question}")
for m in self.recall("failure root cause lesson engineer verdict false alarm OLTC", limit=4):
    if m not in memories:
        memories.append(m)

That second search is what lets the agent connect one transformer's story with another's. The Hindsight documentation and Vectorize's guide on what agent memory is explain the approach in more detail.

An engineer can simply ask:

"Show the DGA history of T-3 and the changes in acetylene."
"Was there any maintenance before the latest increase?"

and get an answer built from the real records.

A Practical Example: T-1 and T-3

T-1 has a record of acetylene appearing a few weeks before a documented failure in 2024.

T-3 has just shown acetylene rising from 0 to 4 ppm after tap-changer maintenance.

Without memory, our agent gave a generic answer: medium risk, check again next month.

With Hindsight memory, it said:

"T-3's acetylene increased from 0 to 4 ppm after last month's tap-changer maintenance. A similar pattern was recorded on T-1 before its documented failure. This similarity does not establish an identical fault, but it is a strong reason for further investigation."

Every answer also shows the exact memories used, so the engineer can check the full DGA report, operating conditions and maintenance records before deciding the next step.

Same AI model. Same reading. The only difference was memory.

How This Improves Maintenance Work

🔍 Less manual searching: ask a question instead of digging through files.

📈 Clearer trends: changes stand out when past readings are right there.

🛠 Maintenance context: test results are seen alongside the work done before them.

🧠 Knowledge that stays: history remains available even when engineers or teams change.

🗂 Organised assets: every transformer gets a structured story, not scattered documents.

Challenges in Building a Reliable Historian

Keeping history sounds simple, but doing it accurately takes care:

Data consistency: labs and teams use different formats, units and transformer IDs.
Missing records: the agent must clearly say when information isn't available.
Correct linking: every event must be tied to the right transformer and date. We put the date inside each memory so the AI always sees the correct order of events.
Careful interpretation: a similar pattern does not mean an identical fault.

What's Next

Interactive trend graphs for each gas.
Automatic extraction of values from lab reports.
Notifications for unusual changes.
Transformer comparison tools and, with enough validated data, anomaly detection.

Any of these would need careful testing, engineering validation and safety controls before real-world use.

Conclusion

A transformer isn't defined by one test report. Its condition develops over time through loading, electrical stress, maintenance and insulation ageing.

Understanding that takes history, and history only helps if it's easy to reach.

With Transformer Health Historian, we connected DGA reports, loading events and maintenance records through Hindsight memory, so engineers can see the context behind every new measurement.

**Every transformer has a story. Our goal is to make that story easy

DEV Community: Prajwal Bevoor